Resumen de Slam in large environments with wearable sensors
In order to perform any valuable task, a really autonomous robot has to be able to localize itself in an unknown environment using its onboard sensors. Simultaneous Localization and Mapping (SLAM) techniques deal precisely with this problem. SLAM algorithms have been implemented for indoor and outdoor mobile robots, unmanned underwater vehicles (UUV) and unmanned aerial vehicles (UAV) using a great variety of sensors such as encoders, lasers, sonars and cameras. However, a difficulty that these algorithms have to face is the scaling problem: when the size of the environment, where the robot performs its tasks, starts growing, the computational requirements of the algorithms increase as well, preventing the robot from working in real time. In addition, when the environment gets larger the global consistency and accuracy of the estimation obtained decrease.
The main objective of the thesis is to develop efficient SLAM algorithms for large environments based on portable and cheap sensors that can be easily carried by a person (wearable sensors).
The type of sensors we consider are: cameras, both monocular and stereo, inertial measurement units (IMUs) and magnetic sensors.
The first main contribution of this thesis is a novel and powerful algorithm in SLAM that allows us to build large maps by decomposing the state into a set of conditionally independent submaps (CI-submaps). As each submap is of constant size, building a submap requires constant time O(1). The key insight of the technique is the conditional independence property (CI-property) between submaps that, in contrast to other submapping approaches, allows submaps to share information consistently. By the CI-property we demonstrate that the global map can be updated and recovered in O(n) without performing any approximation when submaps form simple topologies. That is, we can obtain exactly the same solution as the classical EKF-SLAM algorithm with cost O(1) in local areas and O(n) after a loop closure, instead of the O(n^2) cost per step of EKF-SLAM. Furthermore, when local coordinates are used to represent CI-submaps the solution obtained has even better consistency properties than the EKF, because linearization errors are reduced. In order to use CI-submaps in more complex map topologies we have developed a new algorithm called CI-Graph that extends the computational and consistency properties of CI-submaps without loosing its conditional independence property. Therefore CI-Graph can be classified among the best SLAM algorithms for its computational properties with the additional advantage that no approximations are introduced in the estimation.
The second main contribution of the thesis is a new technique for visual SLAM based on CI-submaps that allows us to build accurate maps of large environments with just a camera in hand either monocular or stereo. CI-submaps are specially well-suited for visual SLAM algorithms since salient features or vehicle states components such as velocity or global attitude can be shared between submaps in a consistent manner. This is extremely valuable to reduce the errors made during the first steps of map initialization and to reduce 'scale-drift' between submaps (specially for monocular vision). For stereo vision, a novel algorithm to incorporate information from features nearby and far from the camera is introduced. Using close points provides scale information through the stereo baseline avoiding 'scale-drift', while inverse depth points are useful to obtain angular information from distant points. Real experiments using a monocular or a stereo camera in hand in indoors and outdoors large environments have been performed. In both cases, even with moving objects in the scene, we have obtained very successful results. Up to the authors' knowledge, this is the first time that such big maps have been achieved using simply a camera in hand.
The third contribution of this thesis is the investigation of the use of new types of wearable sensors in SLAM. First we analyze the advantages of merging information from a monocular camera with IMU readings (Inertial Measurement Unit). As a result of fusing both sensors, the scale of the map can be recovered obtaining better accuracy in the solution even for singular situations. The resultant algorithm has been tested in an outdoor experiment with successful results. Second, we present a novel study of the use of magnetic sensors for SLAM. In particular, we have focused on a rescue application where victims of avalanches have to be localized from magnetic pulses cast by a magnetic device called ARVA. The main difficulty with this transmitter is that the probability distributions obtained during the estimation are far from being Gaussian. We have addressed this problem using an estimation filter based on SOGs (sum of Gaussians) which is not frequently used in the robotics community. Our preliminary results with simulations are quite promising.
We believe that the new theoretical insights and algorithms introduced in this thesis, CI-submaps, CI-Graph and SOGs Filter will open new ways of research for future algorithms. In addition, for visual SLAM, the successful results obtained in large environments with cheap and wearable sensors such as the IMU, monocular and stereo cameras open important future SLAM applications.
